Import Sourcing of Chinese Cities: Order versus Randomness

Transcription

1 Import Sourcing of Chinese Cities: Order versus Randomness Keith Head Ran Jing John Ries First version: May 1, 2010 This version: March 17, 2011 Abstract We utilize very detailed import information for Chinese cities to assess the empirical validity of prominent trade models. Key predictions of the Ricardian comparative advantage and Armington love of variety models are contradicted in the data. We show that cities within a province often do not purchase a narrowly defined product from the leading foreign source of that product in the province. A model of random sourcing goes part way in explaining the observed 64% hierarchy compliance rate. Importing firms orientations towards particular source countries also appear to have a significant impact on the sourcing decisions of cities. This research has been supported by the Social Sciences and Humanities Research Council of Canada. We thank Jacques Thisse, seminar participants at the University of Oxford, UC Davis, and the North American Meetings of the Regional Science Association International for helpful suggestions. Sauder School of Business, University of British Columbia and CEPR, keith.head@sauder.ubc.ca School of International Trade and Economics, University of International Business and Economics, ran.jing@uibe.edu.cn Sauder School of Business, University of British Columbia, john.ries@sauder.ubc.ca

2 1 Introduction Standard trade models make stark predictions about the number of countries from which a country should source goods. Ricardian comparative advantage implies that the low-cost supplier of a product should be the only supplier to that market (single sourcing). On the other hand, in models that assume a love for country-specific varieties (the Armington assumption), and only variable exporting costs, we expect each market to buy from all sources (universal sourcing). More recent models incorporating love of variety, heterogenous firms, and fixed costs predict sourcing from a subset of supplying countries (partial sourcing). This paper investigates the extent to which the sourcing decisions of Chinese cities correspond to predictions of different models of trade. In addition to identifying the predictions of the Ricardian comparative advantage and love of variety trade models, we develop a random model and calculate an expected sourcing outcome. We use very detailed import data for China cities to test model predictions. We document that the incidence of single sourcing by cities is rare, even for goods classified as homogeneous. Universal sourcing is also extremely uncommon, even for differentiated goods. We assume that cities within Chinese provinces are similar geographically relative to supplying countries and, therefore, expect them to have equal access to different exporters. Our analysis establishes that prominent heterogeneous firm models that lead to partial sourcing predict a hierarchy of export sources: All cities that import from a given source country will also buy from the more efficient (lower cost) sources available in the province. An implication of the hierarchy is that all cities source from the most efficient source. Ricardian comparative advantage also predicts sourcing from the lowest cost country. In our random model, the country revealed to host the lowest cost firm is the most likely to be randomly selected but with a probability less than one. The data strongly reject the proposition that all cities source from a most efficient supplying country, thereby providing support for the random sourcing model. However, we also show that cities have different orientations towards specific sources that lead to systematic deviations from the predictions of the random model. We provide evidence indicating that city orientation relates to multinational production networks. Recent research identifies and tests hierarchical predictions generated by heterogeneous-firm models of trade. This literature focuses on the entry of firms (or products) into different destinations. Crozet et al. (2009) adapt the Melitz (2003) model to generate predictions about the quality of 1

3 goods that will be sold in different markets and test the predictions using data on Champagne exports from France. Analyzing heterogeneous foreign direct investors, Chen and Moore (2010) test hierarchical predictions of the Helpman et al. (2004) model and find that only the more productive French firms tend to invest in small or high cost markets. Hierarchy predicts that firms should always ship their strongest products to the destinations where they sell their weaker products but Bernard et al. (2009) find report that this incidence is only 67% for US exporters. Eaton et al. (2008) determine that only 52% of French firms sell to the most popular export market (Belgium). However, additional tests evaluating the sets of export destinations indicate more compliance with hierarchy than random choice of export markets. 1 Our analysis differs in that we focus on the hierarchy of sources for a set of destinations. Observed departures from hierarchy motivate Eaton et al. (2008) and Bernard et al. (2009) to incorporate source-destination effects into trade models with heterogeneous firms. Source-destination effects suggest random sourcing as an alternative to deterministic models. 2 We employ a ballsand-bins approach along the lines of Armenter and Koren (2010). Balls are shipments that randomly fall into bins of different sizes depending on the source country. We find that allowing shipments to be correlated within firms and oriented towards particular source countries provides a better fit to the data. Our paper makes a number of contributions to the literature. First, we measure the extent to which key predictions of the Ricardian comparative advantage model and love of variety models (with and without fixed costs) hold using very disaggregated product and geographic information. We also add to research examining hierarchical predictions of trade models. We propose a simple statistic for evaluating hierarchy compliance and provide a benchmark given by the expected value in a random model. Finally, we document the importance of firm orientation in explaining the pattern of import sourcing. The next section provides theoretical background for the predictions we examine. Section 3 describes the extent of single and universal sourcing of disaggregated products in Chinese cities and how that varies across good types. Partial sourcing is the most prevalent sourcing pattern, a finding consistent with hierarchy models. However, the probability that a city imports from the top provincial source lies well below one, the value predicted by 1 Random entry depends on entry probabilities constructed from overall entry rates. 2 Crozet et al. (2009) also specify a random alternative to hierarchical market entry. 2

4 both the Ricardian comparative advantage and love of variety models. We present a random sourcing model in Section 4 and demonstrate that it provides a reasonable fit to the data. Section 5 reveals that a correlated choice model where firms in cities are oriented towards certain source countries can improve the fit of the random model. The final section summarizes the results and discusses their implications. 2 Theoretical background Two main approaches have guided theoretical and empirical research in international trade in recent years. The first, based on Ricardian comparative advantage, has been generalized by Eaton and Kortum (2002) to a multicountry, multi-product setting that incorporates trade costs. The second approach, introduced by Krugman (1979) stipulates that consumers love variety. In Krugman s models, varieties are associated with firms rather than countries. Feenstra (1994) helped launch a literature reintroducing the Armington (1969) assumption that consumers view products from different countries as different varieties. Paraphrasing Broda and Weinstein (2006), sparkling wine is a product, but sparkling wine from France (i.e. Champagne) is a variety. Given their love of variety, consumers are not satisfied by low prices; they also want to buy the different varieties offered by each supplying country. The more recent heterogeneous firm models of trade assume love of variety and generate partial sourcing and hierarchies of exporters and import destinations. In following subsections we outline the predictions of these prominent models of trade for micro-level sourcing decisions. 2.1 Ricardian comparative advantage In the Ricardian model, generalized to a multi-country model with a continuum of goods by Eaton and Kortum (2002), product characteristics are independent of the country of origin. Hence for each narrowly defined product, a consumer chooses the source country that offers the product at the lowest cost. We specify the delivered cost from source s to destination d as C sd = τ sd c s a s, where τ sd captures trade costs, c s is the cost of a bundle of inputs, and a s is unit factor requirements. Assuming that the low-cost country supplies 3

5 the product elastically, the low cost supplier should take the whole market. Since C sd incorporates transport costs, it is possible that a geographically dispersed country with several neighbors will have the low-cost source for a good vary by region within the importing country. This would result in imports from multiple sources even though each consumption destination within the country buys from a single source country. Panel (a) in Figure 1 portrays three destinations that single source and how this pattern of imports appears to be multisourcing in geographically aggregated data. The key prediction of the Ricardian model is that within narrowly defined geographic regions and products, imports should be observed from a single source country. Our data seem well-suited to testing this prediction because we have products defined at the 8-digit level and our geography is cities within relatively small provinces..&"/$012,%(3//$&/0+&( (a) Single!"#$%&'()%"#*+$,&'-( (b) Universal!"#$%&'()%"#*+$,&'-(./0"$+(1&'2*32"*'()%,2&'-( 4&'5*05"*'( Figure 1: Sourcing in the Ricardian and Armington models 2.2 Armington love of variety Armington (1969) argued that products are distinguished not only by their kind e.g. machinery, chemicals but also by their place of production. Armington went on to specify demands for these national varieties using a constant elasticity of substitution. This functional form implies that no matter what the relative prices are, consumers want to purchase from all available sources, rather than just those from the low-cost supplier. In the case of 4

6 China, most imports are intermediate inputs and capital goods where the importer will usually be the end-user of the imports. In these cases, love-ofvariety is a feature of the production function emphasized by Ethier (1982). The Armington model generates the universal sourcing pattern portrayed in panel (b) in Figure 1. China should buy each good from all the countries that export that good. Within China, cities comprising consumers with Armington preferences should also buy from all countries. However, the amounts consumed from each source depend on the preference parameter for that country s goods as well as the relative prices of each supply country. The basic Armington model has the opposite empirical problem of the Ricardian model. Whereas the Ricardian model predicts single sourcing and there only has a hope of working well on very disaggregated data (narrow product classifications and geographies), the Armington model predicts universal sourcing and therefore can be expected to fit better with highly aggregated data. Haveman and Hummels (2004) point out that even at the SITC4 level of product detail, zero bilateral trade flows are extremely common:...in 99.4% of the cases fewer than half the available varieties [are] purchased...in none of the 75,774 cases does an importer avail itself of all varieties. 2.3 Hierarchies of heterogeneous firms A class of heterogeneous firm models of trade predicts hierarchies of suppliers and destinations in terms of popularity. In these models, the best firms are able to sell profitably in all destinations and the most attractive destinations purchase from all suppliers. These models generally rely upon some form of monopolistic competition in which each supplier offers a differentiated variety. Destination hierarchies exist when a variety sold to the (d + 1)th most popular destination also sells to the dth most popular destination. 3 Source hierarchies occur when a destination that imports from the (s + 1)th most popular source also buys from the sth most popular source. Hierarchies appear in popular heterogeneous-firm trade models in which varieties of goods can be ordered according to single firm-specific variable. That variable could be productivity or quality. It can also be a composite of several underlying factors. Here we model the firm variable as qualityadjusted delivered unit costs to a specific market. The second key ingredient of hierarchy models is that not all varieties are sold in every market. Vari- 3 This definition corresponds to that in Eaton et al. (2008) (page 6). 5

7 eties are not sold if they are priced above the level that chokes off all demand or if the destination market is sufficiently small and/or competitive such that some firms cannot cover the fixed costs of entering the export market. These destination hierarchies obtain because if low-cost varieties that are profitably exported into tough (small, distant, and/or competitive) markets, then they will also be profitably exported into easy (large, nearby, and/or uncompetitive) markets. Similarly, source hierarchies arise because if a variety from one source country is offered to a particular destination, a variety from a lower cost source will also be offered to that destination. In models with heterogenous firms linked to varieties, CES preferences and love of variety, consumers purchase all available varieties but producers only export if operating profits exceed the fixed costs of exporting to a destination. In a continuum of firms (varieties) framework, costs cannot be too low or else there will be firms in every country that profitably export to all destinations and universal sourcing (destinations purchase from all source countries) obtains. Eaton and Kortum (2010) substitute an integer number of firms for the assumptions of a continuum of firms and an upper support for the productivity draw to generate zero trade flows between some countries. Melitz and Ottaviano (2008) do not assume fixed costs but their linear demand model yields hierarchy because marginal costs of some varieties exceed the choke price where demand is zero. Hierarchy will not occur in models where idiosyncratic variety-destination effects cause varieties to be ranked differently by different destinations. We formalize the conditions under which hierarchical relationships are predicted to occur. Our framework involves a precise destination d located within a larger jurisdiction that we denote with upper case D. In the empirical work the d are cities and the D are provinces or province-level municipalities (such as Shanghai). All variables are good-specific but we suppress the good g subscripts for now. Maintaining the notation of Helpman et al. (2008), a i denotes the number of bundles used per unit of output by firm i, c s measures the cost of each input bundle in source country s, and τ sd represents an iceberg form transport cost from source s to destination d. Exporter i from s will therefore have delivered unit costs to market d given by C i sd = τ sdc s a i. We choose units such that c s measures quality-adjusted costs. Thus, differences in the Armington source-country preference parameters are built into C i sd.4 4 Suppose all consumers in all d in region D assign a common preference parameter β sd to each physical unit of production from source s, then Csd i is given by the marginal cost of a physical unit divided by β sd. 6

8 We employ a hub and spoke model for transportation costs from s to d. All goods from s flow to a common point, the hub, in the destination province D and then travel to individual cities via the spokes. The hub could be a large seaport, a regional airport, or a geographical feature such the mouth of the Yangtze river. The key assumption is that no source country has a short cut it can take to reach the final destination. Under a hub and spoke system, the iceberg trade cost factor can be expressed as τ sd = T sd t Dd. 5 Now Csd i can be expressed as the product of the costs of reaching province D s hub (CsD i = c sa i T sd ) and the cost of transporting the good from the hub to city d (t Dd ). The profits of a firm i with unit input requirement a i selling to city d in province D are given by variable profits minus fixed costs, F sd. Variable profits are a function, V (), of delivered unit costs (CsD i t Dd) and a destination demand shifter (Y d ). Thus, profits net of fixed costs are given by π i sd = V (C i sdt Dd, Y d ) F sd, (1) where the partial derivative of the first argument of V () is negative and that of the second argument is positive. Hierarchy predictions require that we solve for a threshold cost level C d such that πsd i < 0 for all i such that Ci sd > C d. This condition implies that a firm i that is good enough to enter market d with C d will also be good enough to enter every other market, d, with C d > C d. To obtain a closed form for C d we employ more of the structure from Helpman et al. (2008). 6. Variable profits are given by λ[csd i t Dd] 1 ɛ Y d P ɛ 1 d, where ɛ is the elasticity of substitution, Y d is expenditure on all varieties, P d is the price index, and λ ɛ ɛ (ɛ 1) ɛ 1. We depart from Helpman et al. (2008) by imposing more structure on the fixed costs of serving each market. In particular, we assume that f d input bundles are required as fixed costs to support positive levels of exporting. 7 Each fixed cost bundle combines inputs from the home country and inputs from the destination market according to a Cobb-Douglas form with share 5 Suppose a fraction δ sd of production melts on the way to the hub, and of the remaining goods, a further faction δ Dd melts along the spoke. To deliver one unit to the final destination therefore requires production of 1/[(1 δ sd )(1 δ Dd )] units. Thus T sd = 1/(1 δ sd ) and t Dd = 1/(1 δ Dd ). 6 While we have not undertaken a full exploration of the necessary conditions for hierarchical sourcing, we believe it would arise in other single-dimension heterogeneous firm models such as Melitz and Ottaviano (2008) 7 Variance in f d across cities might arise due to differences in city size or economic development. 7

9 parameter α. We assume that the same home factor prices, c s, and unit factor requirements, a i, that govern production costs also apply to fixed costs. The destination-level factor costs are denoted w d. To avoid excess notation, we also assume that the cost of supplying factor services from home country s remotely in d is governed by the same trade costs, τ sd, that apply to shipments of goods. Taking these assumptions together we obtain F sd = f d (C i sdt Dd ) α w 1 α d. (2) The key feature of this specification is that fixed costs are multiplicative in a factor that is sd-specific and a factor that is d-specific. It admits the case where fixed costs are independent of s as assumed by Melitz (2003) and Helpman et al. (2004) (when α = 0). Substituting the variable and fixed costs formulas into equation (1) we obtain π i sd = λ[c i sdt Dd ] 1 ɛ Y d P ɛ 1 d f d (C i sdt Dd ) α w 1 α d. (3) We focus on sourcing outcomes of the set of cities in a particular Chinese province and drop the D subscript. Setting equation (3) equal to zero and solving for costs determines the critical level of delivered unit costs, Csd, where profits of serving a particular city d equal zero: C d = 1 t Dd [ λyd P ɛ 1 ] 1 α+ɛ 1 d f d w 1 α d The critical cost level for exporting to city d depends only on d-specific attributes. In particular C d is increasing in the demand shifter Y d and the price index P d but decreasing in local wages and the transport costs from the provincial hub. Based on these characteristics we can order destinations within a province from easiest (highest C d ) to toughest (lowest C d ). The basic idea of hierchical sourcing is that a supplier that is efficient enough to export a tough destination, will export to all easier destinations. We can therefore infer the most efficient supplier by counting the number of markets to which it exports. The key conditions required to generate hierarchy are that the profit function is decreasing in costs, CsD i, the hub and spoke nature of trade costs, and the separable form for fixed costs. 8 The latter two assumptions allow for separating the s-specific and d-specific terms. Without them, the threshold cost for entering a market could depend on s characteristics, leading to 8 The Cobb-Douglas form is not necessary: a fixed cost function that sums s and d terms would also work. 8 (4)

10 breakdown of hierarchical ordering. For example, if a source country had an advantage over its rivals in serving market d but that advantage did not apply to the other markets, then it could export to a market that appeared to be hard but fail to export to easier markets. Lacking data on the individual firms who export to city d we focus on which source countries supply which destination cities. To determine whether source s sells to a city in province D, it is sufficient to focus on whether it is profitable for the most productive firm in s to sell there. For each s, therefore, we define the delivered unit costs to province D s hub of the most profitable (lowest cost) firm in each s as CsD L. Therefore, source s sells to city d if CsD L C d. profit(c,y) C ~ (5) C ~ (10) C ~ (20) C 1 L C 2 L C 3 L C 4 L Delivered Cost (C) Figure 2: Hierarchy Figure 2 depicts hierarchy. The vertical axis shows profits and the horizontal axis shows delivered unit costs C s to a specific province, D. The figure also displays the profit schedules for three cities located in the province with different market conditions, represented by d-specific terms set equal to 5, 10, or 20. The intersection of each profit schedule and the horizontal zero line identifies the critical level of costs that generate zero profits to that particular destination, Cs. The figure also identifies the lowest cost 9

11 firm, Cs L, in source countries 1, 2, 3 and 4. The figure shows that the largest destination imports from all four source countries because C s (20) > Ci L for i=1,2,3,4. Smaller markets import from fewer sources. We observe source hierarchy: All destinations that import from the lowest-cost source country and if a destination imports from the the (s+1)th most popular source, it also sources from the sth most popular source. Destination hierarchy is also evident: If a source finds it profitable to sell to (d+1)th the most popular destination in terms of the number of sources that sell there, it also sells to the dth most popular destination. 3 Empirical evidence The Ricardian comparative advantage model predicts that goods should be single-sourced. The Armington love of variety model predicts universal sourcing whereas heterogeneous-firm models can lead to partial sourcing and hierarchies. We examine the predictions of the models using data on import transactions collected by the Chinese Customs Office for On a monthly basis, we observe each firm s imports by detailed product classification (cn8 level), origin country, port of entry, and destination city in China. 9 Table 1 lists information on China s top 20 imported products according to value. We show the 2006 import value, the number of source countries (#Src), the system of national accounts (SNA) categorization of products (as intermediate, capital, or consumption), the Rauch (1999) classification of differentiated (Dif), reference price (Ref), or organized exchange (Org), and the detailed product description. 10 Eight-digit product classifications are quite detailed: the table shows five separate CN8 categories for integrated circuits. The largest imported product is petroleum and China sourced it from 46 different countries. Indeed, we see no single sourcing of any of the top items and most were sourced from a large number of countries. Exceptions are soy beans and aircrafts between 15 and 45 tons which were sourced from 8 and 4 countries, respectively. Table 1 indicates that the majority of Chinese imports are intermediate goods. Table 2 reveals that the share of intermediates in Chinese imports 9 The harmonized system establishes harmonized classifications out to six digits. Thus, the first six digits in the CN8 correspond to the harmonized system. The last two digits are China-specific classifications. 10 Details on how we attached SNA and Rauch classifications to our data are contained in Appendix A. 10

12 Table 1: Top Products, 2006 CN8 $bil #Src SNA Rauch Description Int Org Petroleum oils (crude) Int Dif Mon. integ. circuits, digital, 0.18 µm Cap Dif Liquid crystal display panels Int Dif Monolithic integrated circuits, not digital Int Dif Mon. int. circ., dig., 0.18 < wid. 0.35µm Int Org Iron ores and concentrates, non-agglomerated Int Dif Mon. integ. circuits, dig., > 0.35µm Dif Fuel oils number Int Org Soya beans, whether or not broken Int Dif Computer parts and accessories Int Dif Hybrid integrated circuits Int Dif Hand-held wireless telephone parts Int Dif Mon. int. circ., dig., 0.18 < wid. 0.35, orig. film Int Dif Terephthalic acid and its salts Cap Aircraft between 15 and 45 tons Cap Dif Computer hard drives Int Ref Copper ores and concentrates Cap Dif Machines and mechanical appliances N.E.S Int Org Cathodes of unwrought copper Int Org Cotton, not carded/combed 11

13 in 2006 was about 75%. The last column of the table is compiled from the Chinese Customs data used in this study. For comparison, we also show the figures using the United Nations Comtrade data base. We observe that information from the two sources closely correspond. The first column shows the breakdown of world exports. Intermediates account for 56% of world exports. Relatively little of Chinese imports are consumption goods 3% compared to 17% for the world. 11 Capital goods account for about 19% of China s imports and 16% of world imports. In our analysis of the sourcing decisions of Chinese, cities, we exclude imports into bonded warehouses. 6.1% of 2006 imports are entrepot and not destined for the Chinese market. Another 4.1% go to other types of bonded warehouses and may not be consumed in the city where the warehouse is located. 12 Excluding this trade, our sample includes 7.9 million monthly shipments of 118,468 firms that import from at least one foreign country. Our primary unit of analysis will be imports of individual cities for specific goods. We have data for 521 cities and 7077 products. The total number of city-product combinations with positive imports is 334, The extent of single and universal sourcing in cities At the national level 95% of the CN8 products and 99.9% of all imports are not single sourced. This suggests that CN8 are differentiated by source country (Armington). Alternatively, China could be too geographically dispersed to be thought of as a single importing entity. Our data is wellsuited to addressing this hypothesis since we observe imports destined to 31 provinces and 521 cities within China. Table 3 provides information on single and universal sourcing of Chinese cities. The unit of observation is a city-good. The last column of the first line of results reveals that imports obtained from just one source country account for almost half these observations. The share is above half for relatively homogeneous goods (Ref and Org). While goods are frequently single-sourced, often a good that is single sourced by one city is not single sourced by other cities: Only 12.9% of all goods are exclusively singlesourced (each city importing the good chooses one source, but not necessarily the same source as other cities that import the good). The total value of 11 On the other hand, the Chinese customs data shows that 31% of Chinese exports are consumption goods. 12 We also exclude observations corresponding to re-imports where the source country was listed as China. 12

14 Table 2: Shares of imports by good type (in %) Comtrade Customs Type World China China Consumption Intermediate Capital Unclassified Table 3: Single and universal sourcing of cities Type of Good: Dif Ref Org All Share (%) of singlesourced imports: City-good observations Goods singlesourced by all cities Value singlesourced relative to total Number of sources per city-cn8: Median Mean Import-weighted avg Share (%) of universally sourced imports: City-good observations with at least 2 sources in the province sources with at least 1% of the province imports both

15 imports accounted for by single sourcing cities is quite small: 6.1% for differentiated goods and 8.5% for the goods sold on organized markets. The next three rows reinforce these results by showing that as we move from medians to simple averages to import-weighted averages that the number of sources per city rises. This tells us that single sourcing is common but there are a small number of cities that source from large numbers of countries and those cities account for a relatively large share of total imports. The Rauch (1999) classification appears to predict the relative amount of single sourcing very successfully. More homogeneous goods are more likely to be single sourced. The problem for the Ricardian model is that if these goods were really homogeneous then they should be exclusively single sourced. The data in the lower half of Table 3 consider the extent of universal sourcing, defined as the share of cities that import from all sources of the good in the province. Overall, 12.9% of cities universally source the goods they import. Surprisingly, universal sourcing is less frequent for differentiated goods. Some of this universal sourcing is a city that imports from the only source in the province (these cities could be the only city that sources the good in the province and only purchase from one source). When we confine the analysis to goods with at least two sources in the province, the incidence of universal sourcing falls by more than half. Some sources in the province may supply a very small amount of imports and, therefore, it is unlikely that all cities will source from them. When we eliminate sources that supply less than 1% of provincial imports, universal sourcing rises to Overall, we find that goods exclusively single sourced are rare, a result that is inconsistent with Ricardian comparative advantage. Even for goods Rauch classifies as homogeneous, only one-quarter of goods are exclusively single sourced. This implies that multisourcing is common. However, a vast majority of cities do not import the same good from all the sources of that available at the province level. The rarity of universal sourcing could be reconciled with the representative consumer love-of-variety model if there were city-specific fixed costs. To test whether such a reconciliation would be supported by the data, we now introduce a statistic to measure the extent that cities comply with the source hierarchy established at the province level. 3.2 A hierarchy statistic We develop a hierarchy statistic to measure the extent that import patterns comply with the hierarchical sourcing prediction of the models. It is calcu- 14

16 lated as the share of cities that import from the top provincial source of the good. The hub-and-spoke assumption implies that under Ricardian comparative advantage, all cities should source from the same low-cost source country. Under the Armington assumption, destinations import from all sources (including the low-cost source). In heterogeneous firm models, a hierarchy emerges where all destinations purchase from the source hosting the producer with the lowest cost. Since we do not have information on which country is the source of the lowest-cost supplier, we must infer it from the data. Table 4 summarizes information on sources of goods for each province. The first column lists the provinces ordered by total imports in 2006, shown in column (2). Guangdong is the largest importer, importing $171 billion. Column (3) and column (4) contain the number of goods imported by the province and the number of cities that import goods. We observe that provinces with more cities tend to import more goods with a higher total value. We identify the top supplying country ( source 1 ) in each province in two ways. First, we calculate source-country shares of provincial imports for each good to identify the top importer of each good in a province. We then deem the country that is most frequently the top source of goods in the province as source 1. Second, for each good we identify the source that is most often chosen by the cities in the province. The country that is most frequently observed as the most frequent source of goods in a province is considered source 1. These rankings may differ because of differences in source-country size. A large source with many firms will sell the highest value of goods but may not necessarily host the lowest cost firm. Column (5) in Table 4 lists source 1 for each province based on highest market share for each good whereas column (7) reveals results based on frequency of being sourced. Japan has is the top source for Guandong in terms of market share for 20.6% of the 6184 goods imported into Guangdong. However, Hong Kong is top in Guangdong based on frequency, being most the most commonly chosen source of Guangdong cities for 25.6% of the products. Across the provinces, the top sources tend to be large traders the United States, Japan, and Germany. We observe some economic geography influencing the choice of top source as Nepal is the top source for Tibet. The top sources are the same in 27 out of 31 provinces across the two identification methods. The few differences that emerge indicate the role of size in the determination of source 1 by value. For example, Japan has the highest market share in Guangdong but the smaller Hong Kong is the top source based on frequency. In Jilin, Germany is supplanted by smaller Korea when frequency is used instead of value. While the top source in 15

17 each province does not change much across the value and frequency methods, there is much more variation for individual goods. Column (8) reveals that for the large importing provinces, the top source is often only the same across the methods about two-thirds or three-quarters of the time. We now calculate the share of cities import a narrowly defined good from the top source of that good in the province. Define y dg as a binary variable equal to 1 if d imports good g from the top source in region D and zero otherwise and let I dg be an indicator that city d imports positive amounts of g from any source. We can express the hierarchy statistic, h 1g, for each good g as the share of all importing cities that source from country 1: h 1g = d y dg. (5) As previously discussed, the source of the highest value of imports may not necessarily host the low-cost firm whereas the source with the lowestcost firm will be most frequently chosen. Thus, the theoretically consistent way to identify the top source country for each good is to see which country is most frequently chosen by cities in a province. In order to have a sufficient number of cities to reliably identify the top source, we impose the restriction that for each good, there must be at least four cities that import the good in the province. 13 This procedure reduces the number of goods from 7077 to The final sample accounts for 82.5% of Chinese imports. Figure 3 presents a histogram of the hierarchy statistic for the 5239 goods. Under the Ricardian comparative advantage, Armington differentiated product, and the heterogeneous firm hierarchy models, the expected value of hierarchy compliance is one. In the Chinese city data we find only 133 goods (2.5% of total) with h 1g = 1. The other 97.5% of the goods do not comply. Both the mean and model compliance is In summary, we observe substantial deviation from the prediction of Ricardian, Armington, and heterogeneous firm hierarchy models that all firms should import from the top source in the province. Eaton et al. (2008) also find widespread departures from hierarchy in their study of French exporters. However, they argue that there is considerably more compliance in the French exporter data than would be predicted by independent choices. In the next section, we evaluate our observed hierarchy statistics relative to a model incorporating an element of randomness in sourcing decision. d I dg 13 In cases where sources are tied for most frequently sourced, we break the tie based on highest import shares. 16

18 Table 4: Top Sources By value By frequency Same Province $mn #(cn8) #(city) Country % Country % % Guangdong Japan 20.6 Hong Kong Jiangsu Japan 29.1 Japan Shanghai Japan 29 Japan Shandong Korea 37.4 Korea Zhejiang Japan 27.3 Japan Beijing Japan 19.2 USA Tianjin Japan 25.2 Japan Liaoning Japan 38.9 Japan Fujian Taiwan 33.5 Taiwan Hebei Japan 21.9 Japan Heilongjiang USA 18.9 USA Hubei Japan 21.6 Japan Jilin Germany 22.8 Korea Anhui Japan 21.1 Japan Sichuan USA 24 USA Henan Japan 21.9 Japan Guangxi Taiwan 15.4 Taiwan Inner Mongolia USA 20.8 USA Yunnan USA 19.8 USA Jiangxi Japan 17.6 Japan Xinjiang USA 30.3 USA Shanxi USA 19.5 Germany Hunan Japan 22.2 Japan Gansu Germany 23.8 Germany Shaanxi USA 24.3 USA Hainan USA 15.8 USA Chongqing Japan 25.5 Japan Guizhou Japan 20.2 Japan Ningxia Germany 28.8 Germany Qinghai Germany 22.8 Germany Tibet Nepal 50 Nepal

19 Mean = 0.64 Counts /4 1/2 3/4 1 Share of cities importing from the province's most frequent source Figure 3: Distribution of hierarchy statistics (h 1 ) for detailed (cn8) goods 4 Random sourcing model A natural way to model non-compliance with hierarchy is to introduce a random element into the choice. A model with randomness does not have to be truly stochastic so long as it contains an idiosyncratic term in the buyer s objective function. The most straightforward way to model this is to employ results from logit random utility models (LRUM) analyzed by Anderson et al. (1992). Individual shipments of goods can be thought of as the smallest potentially independent units of trade. We imagine a procurement agent in each city d that selects the lowest cost source country for each shipment j. The delivered cost perceived by the agent for shipment j in city d is the sum of a common term and an idiosyncratic term. To keep the model as simple as possible, we assume that each country s competitively supplies a homogeneous good with a production cost of c s. The common delivered cost also includes a transport cost, τ sd. In contrast to the hierarchy model where we retained the multiplicative functional forms of Helpman et al. (2008), analytical tractability in the random sourcing model is helped by assuming 18

20 additive transport costs. In this case, the hub and spoke structure sets total transport costs equal to the sum of the costs of reaching the hub, T sd and the transport costs from the hub to the destination, t Dd. Combining production and transport costs, the common delivered cost component for city d in province D from source s is given by c s + T sd + t Dd. The key assumption of the random sourcing model is that individual shipments have idiosyncratic costs, ν j s, associated with each source. We can think of the ν j s as shipment-specific transaction costs, which are presumably the outcome of a prior history of search and experience between the agent firms from each potential source country. 14 The delivered cost inclusive of the random transaction cost is given by C j sd = c s + T sd + t Dd + ν j s. (6) For now, we assume ν j s terms are independent draws from a Gumbel distribution with scale parameter µ. 15 The probability that country s is viewed as lower cost source than any alternative s for any shipment j in city d in province D is given by x sd P[C j sd < Cj s d s s] = exp( (c s + T sd )/µ) h exp( (c h + T hd )/µ). (7) The assumption of additive hub-and-spoke transport costs resulted in the t Dd term dropping out, leaving an expression that lacks any d-specific terms. This because the supplier who has the lowest cost at the hub in province D maintains its advantage when the cost of transporting to the spoke city d is added on. As a result x sd = x sd for all d D. That is, no matter which city in a province a shipment order emanates from, it has the same probability of being filled by a supplier from source country s. The random sourcing model provides microeconomic foundations for the balls-and-bins model that Armenter and Koren (2010) use to explain the incidence of zeros in United States product-country trade flows. We apply the balls-and-bins approach to measure the likelihood that a city will import a good from the top provincial source. Imagine that cities randomly assign (throw) shipments (balls) to source countries (bins). The likelihood that at least one ball from a city will fall into a bin depends on the number 14 In a work-in-progress appendix we will show that the results extend to a model of monopolistic competition between heterogeneous firms. 15 Later, we allow shipments from the same city d to have correlated ν j s, which leads to city orientations towards specific sources. 19

21 of balls and the size of the bins. If n balls are tossed at a bin of size x the probability at least one of them will land in the bin is 1 (1 x) n. For good g, we denote the bin size of the top source in province D as x 1Dg. It is measured as source 1 s share of total shipments of g in province D and reflects the probability shown in equation (7). There is also a cityspecific measured number of shipments, denoted n d. Suppressing the g notation, expected h 1 is d E[h 1 ] = 1 (1 x 1D) n d d I. (8) d This expected value is increasing in both x 1D (the probability a shipment of g to city d in province D will select source 1) and n d (the total number of shipments of good g destined for city d). The hierarchy and random sourcing models are portrayed in Figure 4. There are three import destinations and four source countries. As shown in panel (a), under hierarchy, all destinations import from source 1. Thus, the hierarchy model predicts Eh 1 = 1. Because a market has to be sufficiently large to compensate for the fixed costs of exporting to a market, only the larger two destinations import from source 2. The largest destination is the only one that imports from sources 3 and 4. Panel (b) shows one possible realization of the random sourcing data generating process. The two largest destinations have balls landing in the top country bin (which receives 40% of the balls). However, neither of the balls of the smallest import destination lands in the top country bin. Thus, observed h 1 = 2/3 in this example. Plugging in the x s and n d into (8) we find Eh 1 = (1 (1 0.4) (1 0.4) (1 0.4) 2 )/3 = To implement the random sourcing model, we identify source 1 and its bin size, x 1D. The number of balls thrown towards the bin, n d, are the based on information from each city. This allows us to calculate Eh 1 and compare it to the actual sourcing behaviour. In Table 4, we identified source 1 for each province based on the source with the highest market share for each good as well frequency of city sourcing. We use the latter method here because the former may incorrectly identify large countries as lowest cost. To see this, consider equation (6) in Helpman et al. (2008), M sd = C 1 ɛ sd P ɛ 1 d Y d N s V sd, where V sd is a factor based on the distribution of productivities and N s is the number of firms in s. A high volume of imports, M sd could result from low costs, C sd, peculiarities in the distribution of productivities, V sd, or a 20

22 Figure 4: Sourcing patterns consistent with the same shares (x)!"#$%&'()%"#*+$,&'-( 1&'2*32"*'()%,2&'-(! " #4(567(! $# 4(568(! %# 4(569(!.,/0&*+'( ;,&$3$%.,%3<(!"#$%,*=(! &# 4(56:(!"#$%&'()%"#*+$,&'-( 1&'2*32"*'()%,2&'-(! " #4(567(! $# 4(568(! %# 4(569(!.,/0&*+'( ;3*<"0(!"#$%,*=(! &# 4(56:( 21

23 large number of varieties in the exporting country, N s. In a heterogeneous firm, hierarchy model, all destinations will purchase from the source with the lowest cost firm. This will be the source with the lowest C L, which may not be the source with the greatest volume of exports. By counting the frequency with which cities source positive amounts from each source, we have a popularity rating that will order countries reliably in terms of their least cost suppliers. In order to calculate x 1D and n d, we need to measure shipments. We define a shipment by disaggregating imports by month, country of origin, CN8 good classification, importing firm, route, transport mode, and cityzone. Thus, shipments of good g from source s to city d would be counted separately if they occurred in a different month, were received by a different firm, entered a different port, were routed through different country along the way to China, were transported by a different mode (air, sea, ground), or ended up in a different zone in the city (e.g. Shenzhen SEZ vs Shenzhen city). This measure will be more aggregated than the individual customs declarations used Armenter and Koren (2010) since it lumps together all shipments that occurred in the same month. All together our 2006 data contain 8.4 million shipments (as we define them) compared to 21.6 million customs declarations for the US in The median size of our shipments is $3,278, about twice the $1,800 value in the US data. 16 Given this definition of shipments, we calculate the probability of choosing source 1 under randomness as x 1D = s d n sd d n 1d = n 1D s n. sd We plug x 1D and n d (measured as city shipment) into equation (8) to calculate Eh 1. We then average the values across the provinces to provide an average estimate for the 5239 goods in the sample and compare this expectation to the hierarchy statistic, h 1 generated earlier. Figure 5 presents a scatter plot of the two variables. While Eh 1 appears to be a good predictor of h 1 it is imperfect. Eh 1 tends to be greater than actual h 1 (most balls are to the right of the 45-degree line). Moreover, the slope a regression line is less than one and has an intercept greater than zero. Average Eh 1 across the 5239 goods is 0.71, a higher value than the actual h 1 mean of We showed earlier that intermediates comprise 75% of Chinese imports. Many of these goods are likely to be imported by multinationals who have dedicated relationships to specific source countries, either their home countries or countries where their affiliates reside. Rather than a universal pref- 16 We thank Miklos Koren for providing us the US shipment size data. 22

24 Figure 5: Distribution of hierarchy statistics (h 1 ) for detailed (cn8) goods erence for products of one source country, cities hosting different sets of multinationals may each have different preferences over source countries. Thus, compliance to the hierarchy may be less likely for intermediates. To investigate differences in compliance to the hierarchy or the random model, Table 5 shows average of h 1 and Eh 1 for different subsets of the data based on types of goods. We consider consumption, intermediate and capital goods according to the SNA as well as differentiated, reference, and organized exchange goods as classified by Rauch. We observe that compliance ranges from 0.61 to 0.71, being highest for organized exchange goods and lowest for consumption goods. Across goods, the incidence of noncompliance with the hierarchy is high. The role of multinationals in intermediate goods trade is a possible explanation for low h 1. Multinational companies of different nationalities may locate in different cities and source intermediate from different countries. This would lead to low h 1 and still be consistent with love of variety: Once the factories produce final products they can be shipped to consumers throughout China. The low h 1 for consumption goods seems more damning for the love of variety model. The Chinese customs authority defines 23

25 Table 5: Hierarchy statistics and their expected values Type of good #goods h 1 Eh 1 x 1 #shm All Consumption 1, Intermediate 3, Capital Differentiated 4, Reference Organized The figures in the last four columns are calculated as follows: For h 1, Eh 1, and x 1, we generate the average across importing cities for each good. For #(shm), we calculate the median number of shipments. The table reports the average of these values across the 5239 goods. There are 29 and 134 cn8 categories that we could not assign a SNA and Rauch classification. the destination city as known place within China for consumption, usage, or the final destination of the trip. Since the product from the top provincial source is only imported into 61% of the cities in the case of consumption goods, we can infer that this final good is not available in 39% of the cities. Table 5 shows that Eh 1 generated by the random sourcing model varies across good types. It is relatively high when there is a dominant supplier in the province (large x 1D ) or there are many firms or shipments in cities. The last two columns in the table show average x 1D and the average number of balls. Organized exchange goods have high x 1D and this leads to high values of Eh 1. Comparing actual h 1 to the expectation, we observe that compliance with the hierarchy is always less than what is expected in the random sourcing model. Overall, imports are less likely to be sourced from the source 1 than random occurrence predicts under the assumption that shipments are independent. 5 Cities with source-orientations The random model assumes that the idiosyncratic terms in the random utility model are independent. This assumption implies that cities in a province have common perceptions about the fundamental attractiveness of individual supplying countries and deviations from the common perception are random. We can extend the random model by allowing cities to have orien- 24

26 tation towards particular source countries. We model this orientation as a consequence of heterogeneity in the orientation of importing firms residing in the cities. We can show that cities with orientations towards particular source countries can result in E[h 1 ] being lower than what obtains under independent idiosyncratic terms. Here we focus on a specific province and drop the D subscript. Let the idiosyncratic term associated with shipment j from firm f for source s be denoted νs jf = u f s + e j s. Let e j s be distributed Gumbel with scale parameter µ. Using the expression for trade costs developed in section 4, the cost to firm f located in city d of importing a shipment from source s is Cs jf = c s + T s + t d + νs jf. The probability that g imports from s is x f s P[C jf s < C jf s s s] = exp( (c s + T s + u f s )/µ) h exp( (c (9) h + T h + u f h )/µ). Note that this probability does not depend on which city in the province that the firm is located. We evaluate K cities each with n shipments. The probability that city d purchases the shipment from from source 1 is the shipment-weighted average of the probability that firms located the city purchase from source 1: x 1d = g n f d n xf 1, (10) where n f d is the number of shipments of firm f in city d. Since each city has a 1/K share of shipments in the province, source 1 s share of provincial shipments is x 1 = K x 1d /K. (11) d=1 If u f s = u s, x 1d = x 1 and the expected number of cities that comply is E[h 1 ] iid = K d=1 [1 (1 x 1 ) n ] K = 1 (1 x 1 ) n. (12) In the case of cities with different orientations towards source 1, the expectations is K E[h 1 ] orient [1 (1 x 1d ) n ] =. (13) K d=1 25

27 Let φ(x) = 1 (1 x) n which is a concave function when n > 1. K E[h 1 ] iid = φ(x 1 ) = φ( x 1d /K). (14) d=1 Jensen s inequality implies E[h 1 ] orient = K φ(x 1d )/K. (15) d=1 φ( d α d x d ) d α d φ(x d ). In our application, α d = 1/K. Thus, by Jensen s inequality, E[h 1 ] iid E[h 1 ] orient. City orientation to particular source countries can explain why h 1 is lower than the Eh 1 generated by the IID random sourcing model. Extreme source orientation is depicted in Figure 6. Here cities are oriented towards different source countries. While the shares of source countries, x id, are identical to those in the diagrams portraying hierarchy and random sourcing (Figure 4), the share of cities that comply with the hierarchy is low, 0.33, in contrast to 1 under hierarchy and 0.67 under random with independent balls. Thus, for the same set of x id, we observe very different h1 depending on the model. One source of differences in city orientation may be heterogeneity in the country of origin of foreign-owned firms residing in each city. A city with a high share of firms originating in the province s top source country would be expected to have a higher import orientation towards that country than the province average. We investigate the empirical relevance of differences in city-source orientations by determining the share of firms in each city that are oriented towards source 1. We then use a linear probability model to estimate whether cities that host a higher share of firms oriented towards the province s top source are more likely to comply with the hierarchy by importing from that country. Since the data do not identify the origin-country for firms, we have to infer it using their import patterns. We determine orientation based on each firms s aggregate sourcing patterns across cities and goods. We begin by 26

28 Sources (countries) Destinations (cities) x 1 = 0.4 x 2 = 0.3 x 3 = 0.2 Firms x 4 = 0.1 Figure 6: Source orientation calculating the source country import shares for each firm-city-good combination. The source country with the highest share is deemed to have received a vote in the contest to be the country of orientation for the firm. Then, for each firm, we sum the votes and calculate vote shares for each candidate country. We have a choice of assigning country orientation based on a plurality of votes or a majority of votes. 17 Table 6 shows the results of our assignment. The first two columns reflect outcomes based on the plurality rule whereas the last two columns use the majority rule. We are able to construct orientation for both domestic and foreign firms, the latter comprising both wholly owned and joint ventures. The bottom row reveals that we identify orientations for the majority of foreign firms. We identify orientation for 62.5% of foreign firms under the majority rule and 73.3% under a plurality rule. We are less successful at determining orientation for domestic firms, finding no orientation in 58.8% and 43.2% of the cases under majority and plurality, respectively. The countries to whom Chinese importers are most oriented are Japan, Taiwan, 17 If the total votes equal one, then we consider this firm to have no orientation. In the case of ties for the top spot, we also conclude no orientation. 27